A new method to split acoustic fields in flow-ducts into upstream and downstream propagating sound waves is presented. Goal of this new approach to modal decomposition is to create a methodology that does not rely on semi-analytical solutions of the sound propagation and is, hence, applicable to a much broader set of aeroacoustic problems. An artificial neural network is trained with data gained from a small set of numeric reference solutions of the linearized Navier-Stokes equation in the frequency domain in a straight duct. The network is tested on relevant experimental data. A very good agreement with the analytical solution is demonstrated in two applications with different flow conditions. The good results obtained for the test cases in this paper make this method a promising addition to the classical methodology.